Skip to content
EncryptTechBD
Home Industries Our Work Blog Pricing About Contact Discuss Your Project
বাংলা (BN)
Cloud, DevOps & FinOps 6 min read

Kubernetes operational readiness: five levels compared

Compare Kubernetes readiness levels from a development cluster to a multi-cluster platform, and decide what production actually requires.

E
EncryptTechBD · 0 comments
Woman typing on a laptop
Photo: Burst / StockSnap

Kubernetes is not a maturity badge. It is an operating model with a real cost: upgrades, on-call, security patching, capacity planning and enough platform knowledge to keep the cluster boring. A team that adopts it because a competitor did, without matching the operating model to the workload, usually ends up with more moving parts and less confidence.

Operational readiness should be judged by what the business needs, not by how many features are enabled. This comparison sets out five levels, the criteria that separate them and the point at which each one makes sense.

The five readiness levels

1. Local or development cluster

What it includes: a single-node or desktop cluster for learning and building. No production traffic, no formal backup, no on-call.

Best fit: proof of concept and developer onboarding.

What it does not solve: anything about availability, security or recovery. It should never be described as production-ready.

2. Single managed cluster, basic workloads

What it includes: a managed control plane, one cluster, basic deployments and a shared ingress. Logs and metrics may be minimal.

Best fit: internal tools, low-risk services and teams learning the operating model with vendor support.

What it does not solve: node failure, zone failure, upgrade disruption, secret management or cost control. The blast radius is the whole cluster.

3. Production baseline

What it includes: multiple availability zones, autoscaling node pools, resource requests and limits, health probes, pod disruption budgets, network policies, role-based access, secret management, image scanning, admission policies, backups, central logging, alerting, runbooks and a tested upgrade process.

Best fit: customer-facing services where a few hours of downtime is material but a multi-region active-active design is not justified.

What it does not solve: a bad deployment that passes probes, a regional outage, or a team that cannot respond to alerts. Readiness is a practice, not a checklist.

4. High-availability or multi-cluster

What it includes: clusters in more than one zone or region, a traffic layer that can fail over, replicated state, independent upgrade domains and a tested disaster recovery plan.

Best fit: payment, healthcare, logistics or any service with a contractual recovery objective that a single cluster cannot meet.

What it does not solve: data consistency, split-brain behaviour and the cost of running everything twice. Multi-cluster adds a new class of failure.

5. Internal platform

What it includes: self-service namespaces, golden paths, policy as code, chargeback or showback, standard observability and a platform team that owns the developer experience.

Best fit: many teams shipping frequently, where consistency and autonomy matter more than a single application.

What it does not solve: product priorities. A platform without users is an expensive hobby.

Comparison criteria

  • Workload criticality: internal tool, revenue service or regulated system.
  • Recovery objectives: how long an outage is tolerable and how much data can be lost.
  • Team size and skills: who handles 2am pages, upgrades and security patches.
  • Blast radius: one namespace, one cluster or every customer at once.
  • Compliance: evidence, access review, logging and data residency.
  • Operational burden: the hours per month required to keep the platform healthy.
  • Cost drivers: control plane fees, nodes, load balancers, storage, observability and people.
  • Portability: how much of the design depends on one cloud provider.

Managed versus self-managed

A managed control plane removes some upgrade and availability work, but it does not remove responsibility for workloads, identity, network policy or cost. Self-managed clusters can be appropriate when there is a regulatory reason, an existing platform team or a need for a specific network design. For most small and mid-sized teams, the managed option is the lower-risk starting point, provided the team still learns the workload-level controls. For a Bangladesh-based team, confirm the cloud region, support hours and office connectivity assumptions before committing to a design that assumes always-on access.

Single cluster versus multi-cluster

One cluster with multiple availability zones is the right default for many businesses. It gives zone resilience without the complexity of cross-cluster networking and data replication. Multi-cluster becomes justified when the recovery objective cannot be met inside one cluster, when teams need independent upgrade windows, or when regulation requires separation. Do not adopt it merely to look sophisticated; every additional cluster needs an owner, a backup plan and a test.

The production baseline checklist

  • Named service owner and a written service-level objective.
  • Resource requests and limits for every workload, with a capacity review.
  • Liveness, readiness and startup probes that reflect real behaviour.
  • Pod disruption budgets and topology spread for critical services.
  • Horizontal autoscaling with sensible minimum and maximum replicas.
  • Network policies that default to deny and allow only required traffic.
  • Role-based access, per-environment identities and short-lived credentials.
  • Secrets in a managed store, not in images or Git.
  • Image scanning, signed images and an admission policy.
  • Ingress with TLS, certificate renewal and a documented DNS change process.
  • Backup and restore for persistent data, tested on a schedule.
  • Central logs, metrics, traces and alerts routed to an on-call owner.
  • Runbooks for the top failure modes and a tested cluster upgrade.
  • Cost visibility by namespace or team.
  • CI/CD with rollback and environment separation.

When Kubernetes is the wrong choice

A single application with predictable load, a small team without platform support and a managed application platform that already meets the requirement do not need Kubernetes. A virtual machine with a managed database can be cheaper and easier to operate. The honest test is whether the orchestration benefits, such as scheduling, self-healing and standard deployment patterns, exceed the operating cost for this specific team.

Decision guidance

  • Pilot or learning: level 1. Keep it out of production.
  • Internal tool, one team: level 2, with backups and basic monitoring.
  • Customer-facing service: level 3 minimum. Budget for on-call and upgrades.
  • Contractual recovery target: level 4, with a tested failover.
  • Many product teams: level 5 only after the baseline is stable.

Metrics

Track deployment frequency, change failure rate, time to restore, unscheduled node replacements, upgrade lead time, pod eviction rate, certificate expiry, policy violations and cost per service. A cluster that never fails but takes three weeks to upgrade is not operationally ready.

Where this fits in our work

If you need an independent review of your cluster or a managed operating model, start with Managed Technology & Cybersecurity and review the scope options on our pricing page. For the commercial side of the same decision, read the cloud vendor lock-in guide.

Sources

Services that deliver this

Want this implemented properly?

Pricing →

Every engagement starts with a written scope. If a requirement is outside it, we quote it separately before starting.

Responses (0)

No comments yet. Add the first one below.

Leave a response

Your email address will not be published. Required fields are marked *.

Keep reading

More field notes